PaperlayerView exact revision on arXiv

Source-linked AI summary

CovidGAN: Data Augmentation Using Auxiliary Classifier GAN for Improved Covid-19 Detection

Abdul Waheed, Muskan Goyal, Deepak Gupta, Ashish Khanna, Fadi Al-Turjman, Placido Rogerio Pinheiro

arXiv:2103.05094v1eess.IVcs.CVcs.LG

TL;DR

COVID-19 CXR detection is constrained by the limited radiographic data available for training CNNs. The paper develops CovidGAN, an ACGAN-based generator for synthetic CXR augmentation, and reports that accuracy increased from 85% to 95% when synthetic images were added.

  • Problem

    Limited COVID-19 CXR data makes it difficult to provide CNNs with sufficient training images for detection.

  • Method

    The study develops CovidGAN, an ACGAN-based model that generates synthetic CXR images for augmenting CNN training data.

  • Results

    95% accuracy was achieved with synthetic augmentation, compared with 85% accuracy using CNN classification alone.

  • Takeaways & Limitations

    The authors propose synthetic CXR generation as a way to enhance COVID-19 detection accuracy with minimal data and support stronger radiology systems.

  • Takeaways & Limitations

    The study uses a small multi-source dataset without cross-center validation, and labeling errors could especially affect results when data are scarce.

Abstract

from arXiv · show

Coronavirus (COVID-19) is a viral disease caused by severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2). The spread of COVID-19 seems to have a detrimental effect on the global economy and health. A positive chest X-ray of infected patients is a crucial step in the battle against COVID-19. Early results suggest that abnormalities exist in chest X-rays of patients suggestive of COVID-19. This has led to the introduction of a variety of deep learning systems and studies have shown that the accuracy of COVID-19 patient detection through the use of chest X-rays is strongly optimistic. Deep learning networks like convolutional neural networks (CNNs) need a substantial amount of training data. Because the outbreak is recent, it is difficult to gather a significant number of radiographic images in such a short time. Therefore, in this research, we present a method to generate synthetic chest X-ray (CXR) images by developing an Auxiliary Classifier Generative Adversarial Network (ACGAN) based model called CovidGAN. In addition, we demonstrate that the synthetic images produced from CovidGAN can be utilized to enhance the performance of CNN for COVID-19 detection. Classification using CNN alone yielded 85% accuracy. By adding synthetic images produced by CovidGAN, the accuracy increased to 95%. We hope this method will speed up COVID-19 detection and lead to more robust systems of radiology.

I. INTRODUCTION

COVID-19 detection through chest X-rays is promising, but limited radiographic data makes CNN training difficult. The paper addresses this gap by developing an ACGAN-based approach for synthetic CXR augmentation.

  • Prompt testing and isolation are presented as essential protections because no vaccines or cures existed during the described pandemic context.
  • Chest X-rays can provide visual indicators of coronavirus infection for radiologists, motivating chest-radiography-based detection systems.
  • CNNs achieve strong medical-imaging performance with enough labeled data but can overfit small datasets, while COVID-19 CXR collection remains difficult.
  • Classical augmentation makes slight modifications to existing images, whereas GANs can generate unseen synthetic samples through competing generator and discriminator networks.
  • The study develops an ACGAN-based model to synthesize standard CXR images and expand datasets for COVID-19 detection.
  • Because public COVID-19 CXR images are limited, CovidGAN is proposed to generate artificial training data for CNNs.
  • The paper claims to be the first, to the authors’ knowledge, to present a GAN architecture for improving COVID-19 detection.

B. CONTRIBUTIONS

The study combines a CovidGAN synthetic-image generator with a CNN-based detector to improve COVID-19 classification. Its contributions cover model development, detection architecture, and synthetic augmentation.

  • B. CONTRIBUTIONS: The paper proposes CovidGAN, an ACGAN-based GAN for generating synthetic CXR images.
  • B. CONTRIBUTIONS: It designs a CNN-based model for COVID-19 detection.
  • B. CONTRIBUTIONS: The study augments the CNN training dataset with CovidGAN images to improve COVID-19 detection.
  • B. CONTRIBUTIONS: The paper organizes its dataset and CNN architecture, augmentation method, results, conclusion, and limitations into successive sections.

A. DATASET GENERATION

The dataset combines publicly accessible CXR sources, removes duplicates, and uses VGG16-based CNN processing for COVID-19 detection. The section also describes preprocessing, fine-tuning, and training configuration.

  • A. DATASET GENERATION: 1124 CXR images comprise 403 COVID-CXR and 721 Normal-CXR samples collected from three publicly accessible datasets.
  • A. DATASET GENERATION: Duplicate images are removed after merging the three source datasets using image hashing based on image content.
  • A. DATASET GENERATION: The dataset’s most striking characteristic is the limited public availability of COVID-19 cases and associated CXR images.
  • B. CNN ARCHITECTURE: The CNN uses a VGG16 architecture connected to global average pooling, a 64-unit dense layer, dropout with 0.5 probability, and softmax prediction.
  • B. CNN ARCHITECTURE: Training and testing use 932 and 192 samples respectively, with images resized to 112 × 112 × 3 and normalized before CNN input.
  • B. CNN ARCHITECTURE: Fine-tuning trains custom layers while keeping pretrained VGG16 layers’ weights fixed to adapt learned features to COVID-19 detection.
  • B. CNN ARCHITECTURE: The CNN training setup uses Adam with categorical cross-entropy loss, while the listed implementation details include ReLU activation and 25 epochs.

III. GENERATING SYNTHETIC IMAGES

CovidGAN uses an Auxiliary Classifier GAN to generate synthetic chest X-ray images conditioned on COVID-CXR or Normal-CXR labels, addressing limited medical imaging data for CNN training.

  • III. GENERATING SYNTHETIC IMAGES: Limited medical imaging datasets motivate synthetic augmentation because CNNs can overfit when labeled samples are few and lack variety.Classical augmentation makes only slightly altered versions of existing samples.
  • III. GENERATING SYNTHETIC IMAGES: In a conditional GAN, the generator receives a latent-space point and class label, while the discriminator evaluates whether an image is true or false.AC-GAN instead has the discriminator predict the image’s class label.
  • III. GENERATING SYNTHETIC IMAGES: ACGAN applies class label c and noise z to generate X_fake = G(c, z), while the discriminator outputs probability distributions over sources and class labels.The source and class log-likelihoods form the objective function.
  • III. GENERATING SYNTHETIC IMAGES: CovidGAN is an Auxiliary Classifier GAN that generates synthetic chest X-ray images for COVID-19 detection.The model is based on ACGAN and is proposed specifically to improve COVID-19 detection.
  • III. GENERATING SYNTHETIC IMAGES: The discriminator maximizes L_s + L_c, whereas the generator maximizes L_c - L_s.This objective underlies the proposed CovidGAN architecture.

1) COVIDGAN GENERATOR ARCHITECTURE

The CovidGAN generator combines noise and a class embedding to produce 112 × 112 × 3 images, while the discriminator predicts both image source and CXR class.

  • 1) COVIDGAN GENERATOR ARCHITECTURE: The generator receives normally distributed noise with 0.02 standard deviation and a class label, producing one 112 × 112 × 3 image.The class label passes through a 50-dimensional embedding layer.
  • 1) COVIDGAN GENERATOR ARCHITECTURE: Generator feature maps are upsampled through transpose convolutions from 7 × 7 × 1024 to 14 × 14 × 512, 28 × 28 × 256, 56 × 56 × 128, and 112 × 112 × 3.Noise and class-label tensors are concatenated before upsampling.
  • 1) COVIDGAN GENERATOR ARCHITECTURE: The discriminator is a CNN receiving a 112 × 112 × 3 image and producing real/fake and COVID-CXR/Normal-CXR predictions.Its convolutional blocks include batch normalization, activation, and dropout with 0.5 probability.
  • 1) COVIDGAN GENERATOR ARCHITECTURE: CovidGAN stacks the generator over a non-trainable discriminator so generator updates are driven through the discriminator.It is trained to synthesize both COVID-CXR and Normal-CXR images after resizing and normalizing inputs to [−1, 1].
  • 1) COVIDGAN GENERATOR ARCHITECTURE: CovidGAN generated 1399 synthetic Normal-CXR images and 1669 synthetic COVID-CXR images.Synthetic examples are shown in Figure 4B.

IV. RESULTS AND DISCUSSION

The study evaluates CNN COVID-19 detection with and without synthetic augmentation using accuracy, precision, recall, F1-score, specificity, and macro- and weighted averages.

  • IV. RESULTS AND DISCUSSION: The evaluation uses 192 actual testing samples: 72 COVID-CXR images and 120 Normal-CXR images.The testing set contains only actual data.
  • IV. RESULTS AND DISCUSSION: The study measures CNN performance using precision, recall or sensitivity, F1-score, specificity, and total accuracy.Macro-average and weighted average are also calculated.
  • IV. RESULTS AND DISCUSSION: Precision measures avoiding false positive markings, while recall measures correctly identifying diseased samples.F1-score combines precision and recall, and specificity measures correctly identifying samples without disease.

B. PERFORMANCE ANALYSIS OF SYNTHETIC DATA AUGMENTATION

Synthetic CXR augmentation improved CNN COVID-19 detection performance across accuracy, sensitivity, specificity, precision, and recall compared with actual data alone.

  • B. PERFORMANCE ANALYSIS OF SYNTHETIC DATA AUGMENTATION: 95% accuracy with synthetic augments exceeded 85% accuracy using actual data alone.The corresponding COVID-class precision and recall were 0.96 and 0.90 with synthetic augments, versus 0.89 and 0.69 with actual data.
  • B. PERFORMANCE ANALYSIS OF SYNTHETIC DATA AUGMENTATION: Sensitivity increased from 69% with CNN-AD to 90% with CNN-SA, while specificity increased from 95% to 97%.CNN-AD denotes actual data, whereas CNN-SA combines actual and synthetic images.
  • B. PERFORMANCE ANALYSIS OF SYNTHETIC DATA AUGMENTATION: COVID-class precision and recall reached 0.96 and 0.90, while Normal-class precision and recall reached 0.94 and 0.97.The authors state that the synthetic augments contain meaningful features that help enhance CNN performance.

C. VISUALIZATION USING PCA

The study uses PCA to project CNN features into a lower-dimensional space and compares real and synthetic CXR representations. It also uses confusion matrices to assess COVID-19 classification errors with and without synthetic augmentation.

  • PCA procedure: PCA reduces feature-space dimensionality by creating independent variables while retaining large pairwise distances to optimize variance.The analysis standardizes data, computes covariance and eigenvectors, selects principal components, and projects samples onto the new axes.
  • PCA procedure: The PCA pipeline standardizes dimensions, computes covariance and eigenvectors, selects eigenvectors with the largest eigenvalues, and recasts data along principal-component axes.Standardization ensures each variable contributes equally; selected components define the new subspace for transformed samples.
  • PCA findings: Synthetic-image features lie close to real-image features in the PCA visualization.Features are extracted from the CNN’s last layer, with real and synthetic images plotted together in Figure 5.
  • Confusion matrices: With actual data, the CNN misclassified 22 COVID-CXR images as Normal-CXR and 6 Normal-CXR images as COVID-CXR.The confusion matrix summarizes 192 testing samples, and diagonal entries represent correct classifications.
  • Confusion matrices: With synthetic augmentation, only 7 images were misclassified as Normal-CXR and false positives fell to 3.Figure 7 reports the confusion matrix for CNN training on actual data plus CovidGAN-generated synthetic images.

V. CONCLUSION

The paper concludes that CovidGAN can generate synthetic CXR images to enlarge limited datasets and improve CNN-based COVID-19 detection. It reports higher classification performance with synthetic augmentation and identifies future improvements to synthetic-image quality.

  • V. CONCLUSION: CovidGAN is an ACGAN-based model that generates synthetic CXR images to enlarge the dataset and improve CNN COVID-19 detection.The study used 403 COVID-CXR images and 721 Normal-CXR images.
  • V. CONCLUSION: The CNN was evaluated first on two classes, COVID-CXR and Normal-CXR, and then with synthetic data augmentation.The comparison examines CNN performance with actual data alone versus actual data combined with synthetic images.
  • V. CONCLUSION: 85% to 95% accuracy was recorded when CNN training used actual data and synthetic augments generated by CovidGAN.Increases in precision and recall for both classes were also observed.
  • V. CONCLUSION: Synthesized CXR images showed visualizations and features that help COVID-19 detection.The paper reports a detailed performance analysis of the CNN with synthetic augmentation in Table 1.
  • V. CONCLUSION: The proposed approach enhances COVID-19 detection accuracy with minimal data by generating synthetic chest X-ray images.The authors hope it will support stronger and more reliable radiology systems.
  • V. CONCLUSION: The authors intend to improve synthetic CXR quality by training a Progressive Growing GAN.This is identified as future work following the reported results.

VI. LIMITATIONS

The analysis has limitations involving GAN training, dataset size and provenance, validation, labeling, and the role of clinical testing.

  • VI. LIMITATIONS: The study used a small, multisource dataset because of time constraints and difficulty gathering enough data, and did not conduct cross-center validation.The authors state that more labeled data could improve GAN learning and synthetic-sample quality.
  • VI. LIMITATIONS: Potential labeling mistakes could affect reported results, especially because the dataset is small.The authors report efforts to ensure correct labeling but acknowledge that errors may have a pronounced impact.
  • VI. LIMITATIONS: Reliable COVID-19 detection still requires medical assistance and clinical testing.The paper states that CovidGAN is not intended to compete with laboratory testing.
Loading 2103.05094v1…